Google halved the price of agent work this week, so test your own tasks before you switch

Last Thursday, Google released Gemini 3.7 Flash, which it calls its most capable “workhorse” model for coding and agents. It costs $0.75 per million input tokens and $3.75 per million output tokens through the end of the year, half what Gemini 3.6 Flash cost when it launched. Gemini 3.6 Flash came out 23 days earlier.

If you run a company of 20 to 200 people, you don’t need to follow model releases closely. But this one is a good example of three things you do need to understand: what a task costs, what the model still gets wrong, and how often the ground under you is going to move.

What Google released

Gemini 3.7 Flash is the fast, cheap tier of Google’s lineup. Google aimed it at software work, multi-step agent tasks and document-heavy work in fields like finance and law. It accepts very long inputs, with a context window of one million tokens, enough to load a long stack of contracts in a single request. Output is text only.

Google says it follows instructions more closely, asks for clarification when a request is unclear, and plans its tool calls with more care, which should mean fewer retries. It is available through the Gemini API, through Google’s enterprise agent platform, and inside the Gemini app for paying subscribers through Spark, its always-on personal agent.

The pricing has a catch worth reading twice. The half price is introductory. On January 1, 2027, it doubles to $1.50 and $7.50. Google also moved 3.6 Flash to the same introductory rate, so for now the two cost the same.

Most companies your size won’t call this model directly. You will meet it inside the software you already buy. Plenty of help desk, CRM and document tools run on a model from Google, OpenAI or Anthropic under the hood, and vendors swap those models when a cheaper or better one ships. That can be good news. It can also mean the tool that drafted your quotes well last month behaves a little differently this month, and nobody told you. Ask your vendors which model each AI feature runs on, and whether they tell customers before they change it.

TimelineTwo Flash models in under a month

1. Late JulyGemini 3.6 Flash ships
2. Aug 133.7 Flash ships at half price
3. Dec 31Introductory pricing ends
4. Jan 1, 2027Price doubles to list rate

What a task costs now

Token prices are hard to picture, so here is a hypothetical. Say your support agent reads a customer email plus the order history and your policy notes, 3,000 tokens in, and writes a 500-token draft reply. At the introductory price, that reply costs $0.0041. At the January price, it costs $0.0083. Ten thousand replies a month comes to $41.25 now and $82.50 next year, before any retries or extra steps.

Those numbers are small, and that’s the point. For most companies your size, the model bill is no longer the reason an agent project succeeds or fails. The cost that matters is the time your people spend setting it up, reviewing its work and fixing what it gets wrong. A model that needs fewer retries saves you more in review time than it does in tokens.

Two cautions. Agents rarely make one call per task. A single reply might involve five or ten calls as the agent looks things up, checks rules and rewrites. And plan your budget on the January price, not the introductory one. If the project only makes sense at half price, it doesn’t make sense.

What it still gets wrong

Google published its own scores, and one of them deserves more attention than the coding numbers. AutomationBench tests real business workflows. Gemini 3.7 Flash scored 30.4%, up from 17.0% for 3.6 Flash. On GDP.pdf, a test of document-heavy professional work, it went from 22.0% to 34.0%.

DataGoogle's business-task scores, 3.6 vs 3.7 Flash

• AutomationBench, 3.6 Flash17%
• AutomationBench, 3.7 Flash30.4%
• GDP.pdf, 3.6 Flash22%
• GDP.pdf, 3.7 Flash34%

That is a big jump in three weeks. It is also a model completing well under half of a benchmark built from business work, by the vendor’s own measure. Benchmarks are made hard on purpose, so 30% doesn’t mean your agent fails 70% of your tasks. It does mean you shouldn’t let any model run a whole workflow end to end without a person checking the output. Use the cheap, fast model for the parts of the job that are narrow and checkable, like drafting, sorting, pulling fields from a document or summarizing a long thread, and keep a person on the decision. Refunds, pricing exceptions, anything that goes to a lawyer or a regulator, and anything a customer will read as a promise should pass through someone with a name and a job title before it leaves the building.

The schedule is the real news

Twenty-three days between models. That pace is what I’d plan around, more than any single release.

If every new model sends your team back to rebuild prompts and retest from scratch, you will spend your year chasing releases. The setup I recommend rests on three habits, and none of them require an engineering team:

• Keep a test set of 20 to 50 real examples from the job, with the answer a good employee would give, and run every new model against it.
• Keep your instructions in plain documents the team owns, not buried in one vendor’s settings screen, so moving to another model is a copy and paste.
• Switch only when the test set says the new model does the job better or for less, and write down what changed.

The last habit matters most. “Google says it’s better” is not a reason to switch your production agent. “It got 46 of our 50 shipping-question emails right, up from 41, and costs the same” is a reason. Keep the scores in a shared sheet so anyone can see why the agent runs on the model it runs on.

The same thinking applies to contracts. If a vendor wants a long commitment, ask how they choose models, how often they switch, and whether their price falls when their model costs fall. If their tool runs on a model whose price just dropped by half, they have room to talk. Avoid any agreement that locks a workflow to one named model for a year or more. At this pace, the model in the contract will be several generations old before the term is up.

What to do this week

1. Pick the one agent or AI workflow your company relies on most. Pull 20 real examples of its inputs from last month, and write down what a good answer looks like for each.
2. Run those 20 through your current model and through Gemini 3.7 Flash. Score them yourself or have the person who owns the work do it.
3. Work out the cost per task at the January 2027 price, counting every call the agent makes along the way.
4. Move your instructions into a document your team controls if they currently live only inside a vendor’s tool.
5. Put a recurring 30-minute slot on the calendar, once a month, to rerun the test set against whatever has shipped since.

Sources

• Introducing Gemini 3.7 Flash (Google, August 13, 2026)
• Google releases Gemini 3.7 Flash (DataNorth AI, August 14, 2026)
• Gemini 3.7 Flash (DataCamp, August 14, 2026)